Configure Caffe's AlexNet prototxt files for fine tuning

[BradNeuberg]

Configure Caffe AlexNet prototxt files for fine tuning/transfer learning (freeze some layers, slow down others, etc.). Make sure that the feature scaling values are correct for what we copied over; also configure the solver file to ensure its values are consistent with what we want.

[jhauswald]

Not sure if this goes into this specific issue but given our current setup I should be fairly simple to try two approaches:

• finetuning AlexNet for cloudless
• training from scratch

Just mentioning it so that once our pipeline is nicely streamlined, we can mess with getting better results.

[BradNeuberg]

Now that I'm done prepping the data I'm going to move onto this (and also #3).

[BradNeuberg]

This is done but it's not generalizing terribly well. Once I get EC2 GPU instances setup I'll explore this again or perhaps shift over to VGG-128 to match what Planet Labs is doing.